def test_distance(self):
first = np.array([1, 0, 1, 0, 1, 0, 1], dtype=int)
second = np.array([1, 1, 1, 0, 0, 0, 0], dtype=int)
t1 = TernaryString.fromArray(first)
t2 = TernaryString.fromArray(second)
self.assertEqual(t1.distance(t2), t2.distance(t1))
self.assertEqual(abs(first - second).sum(), t1.distance(t2))
def test_distance(self):
first = np.array([1,0,1,0,1,0,1], dtype=int)
second = np.array([1,1,1,0,0,0,0], dtype=int)
t1 = TernaryString.fromArray(first)
t2 = TernaryString.fromArray(second)
self.assertEqual(t1.distance(t2), t2.distance(t1))
self.assertEqual(abs(first - second).sum(), t1.distance(t2))
def __call__(self, rbm, size=1, initial=None, useAlternate=False):
sample = []
uniform = BernoulliTernary(rbm)
if self.weights is None:
wsample = uniform.batch(10000)
self.weights = []
base = array([exp(-rbm.energy(x, useAlternate)) for x in wsample])
for temp in self.temps:
self.weights.append(((base**(1. / temp)) / 10000.).sum())
for n in xrange(self.chains):
# initialize the chain
x = random.random_sample(rbm.vsize + rbm.hsize)
idx = 1
for l in xrange(self.burn + size / self.chains):
print "ST: chain", n, l
if self.binary:
yv = random.binomial(1, x[:rbm.vsize], rbm.vsize)
else:
yv = x[:rbm.vsize]
yh = dot(yv, rbm.w) + dot(rbm.lh, x[rbm.vsize:]) + rbm.bh
yh = 1. / (1. + exp(-yh / self.temps[idx]))
yh = random.binomial(1, yh, rbm.hsize)
yv = dot(rbm.w, yh) + dot(rbm.lv, yv) + rbm.bv
yv = 1. / (1. + exp(-yv / self.temps[idx]))
if self.binary:
yv = random.binomial(1, yv, rbm.vsize)
y = append(yv, yh, axis=0)
ratio = exp(
(rbm.energy(TernaryString.fromArray(x), useAlternate) -
rbm.energy(TernaryString.fromArray(y), useAlternate)) /
self.temps[idx])
if random.random_sample() < ratio:
x = y
ex = -rbm.energy(TernaryString.fromArray(x))
if random.random_sample() < self.pdown[idx]:
idx2 = idx - 1
p1 = 1 - self.pdown[idx2]
p2 = self.pdown[idx]
else:
idx2 = idx + 1
p1 = self.pdown[idx2]
p2 = 1. - self.pdown[idx]
e1 = exp(ex / self.temps[idx]) * p1 * self.weights[idx]
e2 = exp(ex / self.temps[idx2]) * p2 * self.weights[idx2]
if random.random_sample() < e2 / e1:
idx = idx2
if l >= self.burn:
if self.binary:
sample.append(TernaryString.fromArray(x))
else:
sample.append(x)
return sample
def __call__(self, rbm, size=1, initial = None, useAlternate=False):
sample = []
uniform = BernoulliTernary(rbm)
if self.weights is None:
wsample = uniform.batch(10000)
self.weights = []
base = array([exp(-rbm.energy(x, useAlternate)) for x in wsample])
for temp in self.temps:
self.weights.append(((base ** (1. / temp)) / 10000.).sum())
for n in xrange(self.chains):
# initialize the chain
x = random.random_sample(rbm.vsize + rbm.hsize)
idx = 1
for l in xrange(self.burn + size / self.chains):
print "ST: chain", n, l
if self.binary:
yv = random.binomial(1, x[:rbm.vsize], rbm.vsize)
else:
yv = x[:rbm.vsize]
yh = dot(yv, rbm.w) + dot(rbm.lh, x[rbm.vsize:]) + rbm.bh
yh = 1. / (1. + exp(-yh / self.temps[idx]))
yh = random.binomial(1, yh, rbm.hsize)
yv = dot(rbm.w, yh) + dot(rbm.lv, yv) +rbm.bv
yv = 1. / (1. + exp(-yv / self.temps[idx]))
if self.binary:
yv = random.binomial(1, yv, rbm.vsize)
y = append(yv, yh, axis=0)
ratio = exp((rbm.energy(TernaryString.fromArray(x), useAlternate)-rbm.energy(TernaryString.fromArray(y),useAlternate)) / self.temps[idx])
if random.random_sample() < ratio:
x = y
ex = -rbm.energy(TernaryString.fromArray(x))
if random.random_sample() < self.pdown[idx]:
idx2 = idx - 1
p1 = 1 - self.pdown[idx2]
p2 = self.pdown[idx]
else:
idx2 = idx + 1
p1 = self.pdown[idx2]
p2 = 1. - self.pdown[idx]
e1 = exp(ex / self.temps[idx]) * p1 * self.weights[idx]
e2 = exp(ex / self.temps[idx2]) * p2 * self.weights[idx2]
if random.random_sample() < e2 / e1:
idx = idx2
if l >= self.burn:
if self.binary:
sample.append(TernaryString.fromArray(x))
else:
sample.append(x)
return sample
def mutate(self, x):
"""This method uses an iterative algorithm to flip bits in
a TernaryString with a given probability. The algorithm essentially
uses the binary representation of the bit-flipping probability in order
to convert random byte sampling (``p=.5``) to account for any probability,
with resolution of ``1e-16``.
"""
flipped = TernaryString.bernoulli(self.p(), self.config.space.dim)
base = x.base & ~flipped.base | ~x.base & flipped.base
known = x.known & flipped.known
return TernaryString(base, known, self.config.space.dim)
def __call__(self, rbm, size=1, initial=None, useAlternate=False):
sample = []
for n in xrange(self.chains):
# initialize the chain
x = random.random_sample((len(self.temps), rbm.vsize + rbm.hsize))
idx = 1
for l in xrange(self.burn + size / self.chains):
print "PT: chain", n, l
y = x.copy()
ex = zeros(len(self.temps))
for i, t in enumerate(self.temps):
if self.binary:
yv = random.binomial(1, x[i, :rbm.vsize], rbm.vsize)
else:
yv = x[i, :rbm.vsize]
yh = dot(yv, rbm.w) + dot(rbm.lh, x[i,
rbm.vsize:]) + rbm.bh
yh = 1. / (1. + exp(-yh / t))
yh = random.binomial(1, yh, rbm.hsize)
yv = dot(rbm.w, yh) + dot(rbm.lv, yv) + rbm.bv
yv = 1. / (1. + exp(-yv / t))
if self.binary:
yv = random.binomial(1, yv, rbm.vsize)
y[i] = append(yv, yh, axis=0)
ratio = exp((rbm.energy(TernaryString.fromArray(
x[i]), useAlternate) - rbm.energy(
TernaryString.fromArray(y[i]), useAlternate)) /
self.temps[idx])
if random.random_sample() < ratio:
x[i] = y[i]
ex[i] = rbm.energy(TernaryString.fromArray(x[i]))
#shift temps
for j in xrange(len(self.temps) - 1):
j1 = j
j2 = j + 1
e1 = ex[j1]
e2 = ex[j2]
ratio = exp((e1 - e2) *
(1. / self.temps[j1] - 1. / self.temps[j2]))
if random.random_sample() < ratio:
z = x[j1]
x[j1] = x[j2]
x[j2] = z
if l >= self.burn:
if self.binary:
sample.append(TernaryString.fromArray(x[0]))
else:
sample.append(x[0])
return sample
def test_arrays(self):
first = np.array([1,0,1,0,1,0,1], dtype=int)
second = np.array([1,1,1,0,0,0,0], dtype=int)
t1 = TernaryString.fromArray(first)
t2 = TernaryString.fromArray(second)
self.assertEqual(t1.known, 127L)
self.assertEqual(t1.base, 85L)
self.assertEqual(t1.length, 7)
self.assertEqual(t2.known, 127L)
self.assertEqual(t2.base, 7L)
self.assertEqual(t2.length, 7)
self.assertTrue(np.abs(first - t1.toArray()).max() < 0.01)
self.assertTrue(np.abs(second - t2.toArray()).max() < 0.01)
def test_arrays(self):
first = np.array([1, 0, 1, 0, 1, 0, 1], dtype=int)
second = np.array([1, 1, 1, 0, 0, 0, 0], dtype=int)
t1 = TernaryString.fromArray(first)
t2 = TernaryString.fromArray(second)
self.assertEqual(t1.known, 127L)
self.assertEqual(t1.base, 85L)
self.assertEqual(t1.length, 7)
self.assertEqual(t2.known, 127L)
self.assertEqual(t2.base, 7L)
self.assertEqual(t2.length, 7)
self.assertTrue(np.abs(first - t1.toArray()).max() < 0.01)
self.assertTrue(np.abs(second - t2.toArray()).max() < 0.01)
def __call__(self, rbm, size=1, initial = None, useAlternate=False):
sample = []
for n in xrange(self.chains):
# initialize the chain
x = random.random_sample((len(self.temps), rbm.vsize + rbm.hsize))
idx = 1
for l in xrange(self.burn + size / self.chains):
print "PT: chain", n, l
y = x.copy()
ex = zeros(len(self.temps))
for i, t in enumerate(self.temps):
if self.binary:
yv = random.binomial(1, x[i,:rbm.vsize], rbm.vsize)
else:
yv = x[i,:rbm.vsize]
yh = dot(yv, rbm.w) + dot(rbm.lh, x[i,rbm.vsize:]) + rbm.bh
yh = 1. / (1. + exp(-yh / t))
yh = random.binomial(1, yh, rbm.hsize)
yv = dot(rbm.w, yh) + dot(rbm.lv, yv) +rbm.bv
yv = 1. / (1. + exp(-yv / t))
if self.binary:
yv = random.binomial(1, yv, rbm.vsize)
y[i] = append(yv, yh, axis=0)
ratio = exp((rbm.energy(TernaryString.fromArray(x[i]), useAlternate)-rbm.energy(TernaryString.fromArray(y[i]),useAlternate)) / self.temps[idx])
if random.random_sample() < ratio:
x[i] = y[i]
ex[i] = rbm.energy(TernaryString.fromArray(x[i]))
#shift temps
for j in xrange(len(self.temps) - 1):
j1 = j
j2 = j+1
e1 = ex[j1]
e2 = ex[j2]
ratio = exp((e1 - e2) * (1./self.temps[j1] - 1./self.temps[j2]))
if random.random_sample() < ratio:
z = x[j1]
x[j1] = x[j2]
x[j2] = z
if l >= self.burn:
if self.binary:
sample.append(TernaryString.fromArray(x[0]))
else:
sample.append(x[0])
return sample
def __call__(self, orgs, prob=1.0):
if np.random.random_sample() > prob:
return orgs[0], orgs[1]
if isinstance(orgs[0], np.ndarray):
idx = np.random.random_integers(0, len(orgs[0]) - 1)
return (np.append(orgs[0][:idx], orgs[1][idx:], axis=0),
np.append(orgs[1][:idx], orgs[0][idx:], axis=0))
elif isinstance(orgs[0], TernaryString):
idx = np.random.random_integers(0, orgs[0].length - 1)
ret1 = TernaryString(orgs[0].base, orgs[0].known, orgs[0].length)
ret2 = TernaryString(orgs[1].base, orgs[1].known, orgs[1].length)
ret1[idx:] = orgs[1][idx:]
ret2[idx:] = orgs[0][idx:]
return ret1, ret2
else:
raise ValueError("Received unknown type in OnePointDualCrosser.")
def random(self):
"""Get a random point in binary space. Use the constraint to
generate a random point in the space if possible, otherwise
use a random byte string.
"""
return TernaryString.random(self.dim)
def complete(self, vs, dummy=False):
hs = self.meanFieldUp(len(vs), vs)
xs = [append(vs[i], h, axis=0) for i, h in enumerate(hs)]
return [
TernaryString.fromArray(random.binomial(1, x, shape(x)))
for x in xs
]
def random(self):
"""Get a random point in binary space. Use the constraint to
generate a random point in the space if possible, otherwise
use a random byte string.
"""
return TernaryString.random(self.dim)
def complete2(self, data):
completed = []
for v in data:
x = zeros(self.vsize + self.hsize)
x[: self.vsize] = v
h = dot(v, self.wg) + self.bhg
x[self.vsize :] = 1.0 / (1.0 + exp(-h))
completed.append(TernaryString.fromArray(x))
return completed
def complete2(self, data):
completed = []
for v in data:
x = zeros(self.vsize + self.hsize)
x[:self.vsize] = v
h = dot(v, self.wg) + self.bhg
x[self.vsize:] = 1. / (1. + exp(-h))
completed.append(TernaryString.fromArray(x))
return completed
def __call__(self,
rbm,
size=1,
initial=None,
useAlternate=False,
clampIdx=None):
from pyec.util.TernaryString import TernaryString
sample = []
vsize = rbm.vsize
hsize = rbm.hsize
if useAlternate:
w = rbm.wg
bv = rbm.bvg
bh = rbm.bhg
else:
w = rbm.w
bv = rbm.bv
bh = rbm.bh
if clampIdx is None:
clampIdx = 0
if initial is not None:
initialArr = initial.toArray(rbm.dim)
for i in xrange(self.chains):
if initial is not None:
x = initialArr
else:
T0 = 1. / 4
x = random.random_sample(vsize + hsize).round()
h = dot(x[:vsize], w) + bh
h = 1. / (1. + exp(-h * T0))
if self.meanField:
x[vsize:] = h
else:
x[vsize:] = random.binomial(1, h, hsize)
for j in xrange(self.burn + size / self.chains):
print "Gibbs: chain ", i, j
T = maximum((3 * self.burn + j) / float(4 * self.burn), 1.0)
v = dot(w, x[vsize:]) + bv
if self.meanField:
x[clampIdx:vsize] = (1. / (1. + exp(-v * T)))[clampIdx:]
else:
x[clampIdx:vsize] = random.binomial(
1, 1. / (1. + exp(-v * T)), vsize)[clampIdx:]
h = dot(x[:vsize], w) + bh
x[vsize:] = 1. / (1. + exp(-h * T))
if self.binary and not self.meanField:
x[vsize:] = random.binomial(1, x[vsize:], hsize)
if j >= self.burn:
if self.binary:
y = x
if self.meanField:
y = random.binomial(1, x, vsize + hsize)
sample.append(TernaryString.fromArray(y))
else:
sample.append(x)
return sample
def partition(self):
""" compute the partition function - only for small dimension !!! """
from pyec.util.TernaryString import TernaryString
total = 0
vsize = self.vsize
hsize = self.hsize
all = (1L << (vsize + hsize)) - 1L
for i in xrange(1 << (vsize + hsize)):
total += exp(self.__call__(TernaryString(long(i), all)))
return total
def mutate(self, x):
"""This method uses an iterative algorithm to flip bits in
a TernaryString with a given probability. The algorithm essentially
uses the binary representation of the bit-flipping probability in order
to convert random byte sampling (``p=.5``) to account for any probability,
with resolution of ``1e-16``.
"""
flipped = TernaryString.bernoulli(self.p(), self.config.space.dim)
base = x.base & ~flipped.base | ~x.base & flipped.base
known = x.known & flipped.known
return TernaryString(base, known, self.config.space.dim)
def complete(self, data, sample=True):
completed = []
for v in data:
x = zeros(self.vsize + self.hsize)
x[: self.vsize] = v
h = dot(v, self.w) + self.bh
if sample:
x[self.vsize :] = random.binomial(1, 1.0 / (1.0 + exp(-h)), self.hsize)
completed.append(TernaryString.fromArray(x))
else:
x[self.vsize :] = 1.0 / (1.0 + exp(-h))
completed.append(x)
return completed
def complete(self, data, sample=True):
completed = []
for v in data:
x = zeros(self.vsize + self.hsize)
x[:self.vsize] = v
h = dot(v, self.w) + self.bh
if sample:
x[self.vsize:] = random.binomial(1, 1. / (1. + exp(-h)),
self.hsize)
completed.append(TernaryString.fromArray(x))
else:
x[self.vsize:] = 1. / (1. + exp(-h))
completed.append(x)
return completed
def completeV(self, data, sample=True):
completed = []
for h in data:
h2 = h.toArray(self.bh.size)
x = zeros(self.vsize + self.hsize)
x[self.vsize :] = h2
v = dot(self.w, h2) + self.bv
if sample:
x[: self.vsize] = random.binomial(1, 1.0 / (1.0 + exp(-v)), self.vsize)
completed.append(TernaryString.fromArray(x))
else:
x[: self.vsize] = 1.0 / (1.0 + exp(-v))
completed.append(x)
return completed
def __call__(self, rbm, size=1, initial=None, useAlternate=False, clampIdx=None):
from pyec.util.TernaryString import TernaryString
sample = []
vsize = rbm.vsize
hsize = rbm.hsize
if useAlternate:
w = rbm.wg
bv = rbm.bvg
bh = rbm.bhg
else:
w = rbm.w
bv = rbm.bv
bh = rbm.bh
if clampIdx is None:
clampIdx = 0
if initial is not None:
initialArr = initial.toArray(rbm.dim)
for i in xrange(self.chains):
if initial is not None:
x = initialArr
else:
T0 = 1. / 4
x = random.random_sample(vsize+hsize).round()
h = dot(x[:vsize], w) + bh
h = 1. / (1. + exp(-h*T0))
if self.meanField:
x[vsize:] = h
else:
x[vsize:] = random.binomial(1, h, hsize)
for j in xrange(self.burn + size / self.chains):
print "Gibbs: chain ", i, j
T = maximum(( 3*self.burn + j ) / float(4*self.burn), 1.0)
v = dot(w, x[vsize:]) + bv
if self.meanField:
x[clampIdx:vsize] = (1. / (1. + exp(-v*T)))[clampIdx:]
else:
x[clampIdx:vsize] = random.binomial(1, 1. / (1. + exp(-v*T)), vsize)[clampIdx:]
h = dot(x[:vsize], w) + bh
x[vsize:] = 1. / (1. + exp(-h*T))
if self.binary and not self.meanField:
x[vsize:] = random.binomial(1, x[vsize:], hsize)
if j >= self.burn:
if self.binary:
y = x
if self.meanField:
y = random.binomial(1, x, vsize+hsize)
sample.append(TernaryString.fromArray(y))
else:
sample.append(x)
return sample
def __call__(self, network):
numBytes = int(ceil(len(network.variables) / 8.0))
numFull = len(network.variables) / 8
initial = ''
if numBytes != numFull:
extra = len(network.variables) % 8
initMask = 0
for i in xrange(extra):
initMask <<= 1
initMask |= 1
initial = struct.pack('B',initMask)
base = long(binascii.hexlify(random.bytes(numBytes)), 16)
known = long(binascii.hexlify(initial + '\xff'*numFull), 16)
return TernaryString(base, known)
def __call__(self, orgs, prob):
if np.random.random_sample() > prob:
return orgs[0]
if self.config.space.type == np.ndarray:
rnd = np.random.random_sample(len(orgs[0])).round()
return rnd * orgs[0] + (1 - rnd) * orgs[1]
elif self.config.space.type == TernaryString:
rnd = np.random.bytes(len(str(orgs[0])) * 8)
rnd = long(binascii.hexlify(rnd), 16)
base = rnd & orgs[0].base | ~rnd & orgs[1].base
known = rnd & orgs[0].known | ~rnd & orgs[1].known
return TernaryString(base, known, self.config.space.dim)
else:
err = "Unknown type for UniformCrossover: {0}"
raise NotImplementException(err.format(self.config.space.type))
def completeV(self, data, sample=True):
completed = []
for h in data:
h2 = h.toArray(self.bh.size)
x = zeros(self.vsize + self.hsize)
x[self.vsize:] = h2
v = dot(self.w, h2) + self.bv
if sample:
x[:self.vsize] = random.binomial(1, 1. / (1. + exp(-v)),
self.vsize)
completed.append(TernaryString.fromArray(x))
else:
x[:self.vsize] = 1. / (1. + exp(-v))
completed.append(x)
return completed
def complete(self, vs, dummy=False):
hs = self.meanFieldUp(len(vs), vs)
xs = [append(vs[i], h, axis=0) for i,h in enumerate(hs)]
return [TernaryString.fromArray(random.binomial(1, x, shape(x))) for x in xs]
def __call__(self, rbm, size=1, initial=None, useAlternate=False):
from pyec.util.TernaryString import TernaryString
sample = []
vsize = rbm.vsize
hsize = rbm.hsize
if useAlternate:
w = rbm.wg
bv = rbm.bvg
bh = rbm.bhg
else:
w = rbm.w
bv = rbm.bv
bh = rbm.bh
lv = rbm.lv
lh = rbm.lh
biases = append(bv, bh, axis=0)
wp = 1. - exp(-2 * w)
wn = 1. - exp(2 * w)
lvp = 1. - exp(-2 * lv)
lvn = 1. - exp(2 * lv)
lhp = 1. - exp(-2 * lh)
lhn = 1. - exp(2 * lh)
if initial is not None:
initialArr = initial.toArray(rbm.dim)
start = time()
for n in xrange(self.chains):
start = time()
if initial is not None:
x = initialArr
else:
x = random.random_sample(vsize + hsize).round()
h = dot(x[:vsize], w) + bh
x[vsize:] = random.binomial(1, 1. / (1. + exp(-h)), hsize)
last = copy(x)
repeats = 0
for l in xrange(self.burn + size / self.chains):
print "SW: chain ", n, l
startInner = time()
clusters = array([i for i in xrange(vsize + hsize)], dtype=int)
r = random.random_sample((vsize, hsize))
rh = random.random_sample((hsize, hsize))
rv = random.random_sample((vsize, vsize))
for i in xrange(vsize):
for j in xrange(hsize):
i2 = j + vsize
if (w[i,j] < 0.0 and x[i] != x[i2] and r[i,j] < wn[i,j]) or \
(w[i,j] > 0.0 and x[i] == x[i2] and r[i,j] < wp[i,j]):
#if d[i,j] > 0.0:
if clusters[i] != clusters[i2]:
# take the lesser number
if clusters[i] < clusters[i2]:
c1 = clusters[i]
c2 = clusters[i2]
else:
c1 = clusters[i2]
c2 = clusters[i]
ifi2 = maximum(1 - abs(clusters - c2), 0)
noti2 = 1 - ifi2
clusters = ifi2 * c1 + noti2 * clusters
for i in xrange(vsize):
for j in xrange(vsize):
if i == j: continue
i2 = j
if (lv[i,j] < 0.0 and x[i] != x[i2] and rv[i,j] < lvn[i,j]) or \
(lv[i,j] > 0.0 and x[i] == x[i2] and rv[i,j] < lvp[i,j]):
#if d[i,j] > 0.0:
if clusters[i] != clusters[i2]:
# take the lesser number
if clusters[i] < clusters[i2]:
c1 = clusters[i]
c2 = clusters[i2]
else:
c1 = clusters[i2]
c2 = clusters[i]
ifi2 = maximum(1 - abs(clusters - c2), 0)
noti2 = 1 - ifi2
clusters = ifi2 * c1 + noti2 * clusters
for i in xrange(hsize):
for j in xrange(hsize):
if i == j: continue
i2 = j
if (lh[i,j] < 0.0 and x[i] != x[i2] and rh[i,j] < lhn[i,j]) or \
(lh[i,j] > 0.0 and x[i] == x[i2] and rh[i,j] < lhp[i,j]):
#if d[i,j] > 0.0:
if clusters[i] != clusters[i2]:
# take the lesser number
if clusters[i] < clusters[i2]:
c1 = clusters[i]
c2 = clusters[i2]
else:
c1 = clusters[i2]
c2 = clusters[i]
ifi2 = maximum(1 - abs(clusters - c2), 0)
noti2 = 1 - ifi2
clusters = ifi2 * c1 + noti2 * clusters
computed = zeros(x.size)
for k in xrange(x.size):
if computed[k] > 0.0:
continue
incluster = maximum(1 - abs(clusters - clusters[k]), 0)
bias = (incluster * biases).sum()
if self.binary:
state = float(
random.binomial(1, 1. / (1. + exp(-bias)), 1))
else:
state = float(1. / (1. + exp(-bias)))
computed += incluster
x = incluster * state + (1 - incluster) * x
if l >= self.burn:
if self.binary:
sample.append(TernaryString.fromArray(x))
else:
sample.append(x)
x = random.binomial(1, state, x.size)
print "SW sample ", n, ": ", time() - start
return sample
def test_random(self):
total = np.zeros(100)
for i in xrange(10000):
total += TernaryString.random(100).toArray() / 10000.0
print "Total max: ",
self.assertTrue(np.abs(total - .5).max() < .025)
def extent(self):
lower = 0L | (self.spec.known & self.spec.base)
upper = -1L & (self.spec.known & self.spec.base | ~self.spec.known)
return (TernaryString(lower, self.spec.known, self.spec.length),
TernaryString(upper, self.spec.known, self.spec.length))
def values(self):
mask = 1L << self.index
yield TernaryString(0L, mask)
yield TernaryString(mask, mask)
raise StopIteration
def configurations(self):
cnt = 1L << (len(self.parents))
for i in xrange(cnt):
yield TernaryString(self.expand(long(i)), self.known)
raise StopIteration
def test_random(self):
total = np.zeros(100)
for i in xrange(10000):
total += TernaryString.random(100).toArray() / 10000.0
print "Total max: ",
self.assertTrue(np.abs(total - .5).max() < .025)
def __call__(self, rbm, size=1, initial = None, useAlternate=False):
from pyec.util.TernaryString import TernaryString
sample = []
vsize = rbm.vsize
hsize = rbm.hsize
if useAlternate:
w = rbm.wg
bv = rbm.bvg
bh = rbm.bhg
else:
w = rbm.w
bv = rbm.bv
bh = rbm.bh
lv = rbm.lv
lh = rbm.lh
biases = append(bv, bh, axis=0)
wp = 1. - exp(-2 * w )
wn = 1. - exp(2 * w )
lvp = 1. - exp(-2 * lv)
lvn = 1. - exp(2 * lv)
lhp = 1. - exp(-2 * lh)
lhn = 1. - exp(2 * lh)
if initial is not None:
initialArr = initial.toArray(rbm.dim)
start = time()
for n in xrange(self.chains):
start = time()
if initial is not None:
x = initialArr
else:
x = random.random_sample(vsize + hsize).round()
h = dot(x[:vsize], w) + bh
x[vsize:] = random.binomial(1, 1. / (1. + exp(-h)), hsize)
last = copy(x)
repeats = 0
for l in xrange(self.burn + size / self.chains):
print "SW: chain ", n, l
startInner = time()
clusters = array([i for i in xrange(vsize + hsize)], dtype=int)
r = random.random_sample((vsize, hsize))
rh = random.random_sample((hsize, hsize))
rv = random.random_sample((vsize, vsize))
for i in xrange(vsize):
for j in xrange(hsize):
i2 = j + vsize
if (w[i,j] < 0.0 and x[i] != x[i2] and r[i,j] < wn[i,j]) or \
(w[i,j] > 0.0 and x[i] == x[i2] and r[i,j] < wp[i,j]):
#if d[i,j] > 0.0:
if clusters[i] != clusters[i2]:
# take the lesser number
if clusters[i] < clusters[i2]:
c1 = clusters[i]
c2 = clusters[i2]
else:
c1 = clusters[i2]
c2 = clusters[i]
ifi2 = maximum(1 - abs(clusters - c2), 0)
noti2 = 1 - ifi2
clusters = ifi2 * c1 + noti2 * clusters
for i in xrange(vsize):
for j in xrange(vsize):
if i == j: continue
i2 = j
if (lv[i,j] < 0.0 and x[i] != x[i2] and rv[i,j] < lvn[i,j]) or \
(lv[i,j] > 0.0 and x[i] == x[i2] and rv[i,j] < lvp[i,j]):
#if d[i,j] > 0.0:
if clusters[i] != clusters[i2]:
# take the lesser number
if clusters[i] < clusters[i2]:
c1 = clusters[i]
c2 = clusters[i2]
else:
c1 = clusters[i2]
c2 = clusters[i]
ifi2 = maximum(1 - abs(clusters - c2), 0)
noti2 = 1 - ifi2
clusters = ifi2 * c1 + noti2 * clusters
for i in xrange(hsize):
for j in xrange(hsize):
if i == j: continue
i2 = j
if (lh[i,j] < 0.0 and x[i] != x[i2] and rh[i,j] < lhn[i,j]) or \
(lh[i,j] > 0.0 and x[i] == x[i2] and rh[i,j] < lhp[i,j]):
#if d[i,j] > 0.0:
if clusters[i] != clusters[i2]:
# take the lesser number
if clusters[i] < clusters[i2]:
c1 = clusters[i]
c2 = clusters[i2]
else:
c1 = clusters[i2]
c2 = clusters[i]
ifi2 = maximum(1 - abs(clusters - c2), 0)
noti2 = 1 - ifi2
clusters = ifi2 * c1 + noti2 * clusters
computed = zeros(x.size)
for k in xrange(x.size):
if computed[k] > 0.0:
continue
incluster = maximum(1 - abs(clusters - clusters[k]), 0)
bias = (incluster * biases).sum()
if self.binary:
state = float(random.binomial(1, 1. / (1. + exp(-bias)), 1))
else:
state = float(1. / (1. + exp(-bias)))
computed += incluster
x = incluster * state + (1-incluster) * x
if l >= self.burn:
if self.binary:
sample.append(TernaryString.fromArray(x))
else:
sample.append(x)
x = random.binomial(1, state, x.size)
print "SW sample ", n, ": ", time() - start
return sample
